You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Existing Pi contributor and CI uses do not test the supported user onboarding and sandbox lifecycle.
NemoClaw needs release evidence bound to the exact candidate commit, managed image, OpenShell version, inference route, compute runtime, and policy.
Pi may ship before native Podman activates. The test plan must not block Pi on unfinished Podman work unless #7926 includes Podman at launch, but Podman must include Pi before #7744 claims all-shipped-agent support.
Desired Behavior
Deterministic projects cover Pi manifests, managed-image contracts, configuration, onboarding, lifecycle, inference, state, security boundaries, and compute-runtime neutrality.
A fresh live environment on each supported launch platform completes one interactive and one non-interactive tool-using Pi task through the public NemoClaw path.
If Podman is launch scope, the same exact Pi image and agent contracts pass under native Podman. If Podman is deferred, #7744 tracks Pi as required activation evidence.
Qualification contract
Before implementation begins, #7926 must record the platform, architecture, compute-runtime, provider and API-family matrix, model-validation rules, and exact release qualification model.
The qualification plan must define one versioned deterministic tool-using task, an independent success oracle, timeout and retry limits, and a repository-owned evidence manifest location. Agent self-reporting is not a sufficient oracle.
Live Qualification Flow
Install the exact NemoClaw candidate through the supported path.
Onboard Pi through public non-interactive inputs.
Verify recorded agent, compute runtime, managed-image digest, inference route, and policy identities.
Prove that stock onboarding did not build a repository Dockerfile.
Run a deterministic tool-using task through managed inference.
Run an interactive terminal flow.
Rebuild the sandbox and repeat the task.
Verify declared persistent state and denied access.
Repeat under each compute runtime accepted for Pi launch.
Capture logs and bounded evidence without credentials.
Constraints and Non-goals
Do not qualify source patches or manual sandbox-state edits.
Do not use an unpinned runtime, image, or mutable alias.
Do not treat an agent assertion as the only success oracle.
Summary
Qualify the supported Pi runtime with deterministic tests and exact-candidate live E2E evidence across the accepted compute-runtime matrix.
Parent epic: #7923
Coordination epic: #7744
Landed dependency baseline: #7756, #7772, #8234, #8143, and #7718.
Depends on #7926, #7925, #7927, #7930, and #7924. #7929 consumes this evidence, and #8818 owns final activation.
Target completion: 2026-08-25.
Problem Statement
Existing Pi contributor and CI uses do not test the supported user onboarding and sandbox lifecycle.
NemoClaw needs release evidence bound to the exact candidate commit, managed image, OpenShell version, inference route, compute runtime, and policy.
Pi may ship before native Podman activates. The test plan must not block Pi on unfinished Podman work unless #7926 includes Podman at launch, but Podman must include Pi before #7744 claims all-shipped-agent support.
Desired Behavior
Deterministic projects cover Pi manifests, managed-image contracts, configuration, onboarding, lifecycle, inference, state, security boundaries, and compute-runtime neutrality.
A fresh live environment on each supported launch platform completes one interactive and one non-interactive tool-using Pi task through the public NemoClaw path.
If Podman is launch scope, the same exact Pi image and agent contracts pass under native Podman. If Podman is deferred, #7744 tracks Pi as required activation evidence.
Qualification contract
Before implementation begins, #7926 must record the platform, architecture, compute-runtime, provider and API-family matrix, model-validation rules, and exact release qualification model.
The qualification plan must define one versioned deterministic tool-using task, an independent success oracle, timeout and retry limits, and a repository-owned evidence manifest location. Agent self-reporting is not a sufficient oracle.
Live Qualification Flow
Constraints and Non-goals
Acceptance Criteria
Category
Testing
Checklist